Research On Storage Optimization Based On Solid State Drive’s Characteristics

Storage systems which are built on traditional mechanical Hard Disk Drives(HDD) have long suffered from the I/O bottleneck problem in computer systems. Especially, in the context of increasingly expanding of computing power and computing requirements which are for example driven by the emergence of multicore processors and big-data applications, the performance gap between processer and storage system is getting even bigger, aggravat-ing the I/O bottleneck problem. The main reason that is responsible for the I/O bottleneck lies in the access characteristics of HDDs, e.g., non-random access, platter seeking, rotating delay, etc. During the past decades, flash-based Solid State Drives(SSD) technology has improved drastically, with manufacture cost dropping significantly and performance/cost approaching or even better than that of HDD. As a consequence, SSD are gradually entering in storage systems and getting deployed widely. As opposed to HDDs, SSDs are completely composed of semiconduct technology and has no mechanical components. SSDs have a number of superior features, like high-performance, high-reliability, random access, low power consumption, etc. SSD technology has been projected to be able to eliminate the storage system I/O bottleneck and brings fundamental changes to storage systems.However, though SSDs have apparent advantages over HDDs, they have their own characteristics and idiosyncrasies due to their different structural organizations. One of the most outstanding and critical features is that the basic building blocks of SSDs are Electrically Erasable Programmable Read-Only Memory(EEPROM) chips, causing SSD to be inherently "out-of-place update" and have limited number of erase cycles (lifetime). To fully leverage all of the constituent chips as a whole, SSD introduces an intermediate layer called Flash Translation Layer(FTL) internally to perform functionalities like logical-to-physical address mapping, wear-leveling and garbage collection. As semiconduct technol-ogy steadily advances and widespread deployment and growing importance of SSDs in the storage systems, it is important and fruitful to conduct research on how to best utilize SSD in modern storage systems, while avoiding their shortcomings. This thesis has attempted to best utilize SSDs in storage systems and achieve the following results.(1) Taking advantage of the "no in-place update" idiosyncrasy of SSDs, we propose BVSSD, a scheme that preserves those internally superseded but lingering historical data to realize block-level Continuous Data Protection(CDP) functionality. In contrast to mechan-ical magnetic surface recording, SSDs support read, write(program) and erase operations. SSDs’storage chips are inherently a type of highly hierarchical structure, composed of page, block, plane, die and chip. Both read and write operations are page based, while erase oper-ation must be carried out in unit of blocks. Pages much be erased to be free before they can be rewritten. For each incoming write, SSD first allocates a free page and writes the data to the page. After that, it updates the FTL mapping table to reflect the new status. BVSSD keeps track of and preserves the FTL changing history and simply restores the FTL state to a previous timepoint when doing recovery. It is a lightweight implementation of continuous data protection.(2) A new SSD-oriented I/O scheduler named PASS is proposed with the underlying design of leveraging the rich parallelism within SSDs. SSDs are composed of multiple chips and each chip in turn is also composed of multi-layers, making rich parallelism available within SSDs which could potentially be leveraged to improve performance. Parallelism can be inter-channel, inter-chip, inter-die and inter-plane. PASS partitions the whole storage space into fixed size regions as basic scheduling units and associates a dedicated dispatching queue with each of the regions. PASS dispatches the incoming block write requests into their responsible dispatching queues according to their access addresses and adopts techniques to avoid read-write interference. Experimental results have demonstrated that PASS has better utilized the device internal parallelism and has improved performance.(3) A hybrid storage scheme named HSStore which comprises of both Solid State Drives(SSD) and mechanical disks is proposed to leverage the fact that SSD and mechanical disks have different and complementary merits and disadvantages. In HSStore, SSDs act as a cache layer above mechanical disks. The request dispatcher monitors the incoming I/O pattern and redirects both big sequential and small random writes to disk directly, bypassing the SSD cache layer. Furthermore, it also tracks those missing reads and instructs data mi-grator to copy those data blocks having high miss rate from disks to SSD. Thanks to these combining techniques, HSStore would utilize SSD space more effectively and extending its usage life in the mean while.Through the above investigations, this thesis presents comprehensive researches on the deployment of SSD in modern storage systems from different perspectives. It has been demonstrated that SSDs bear great potential in their usage in storage systems and if we can fully take their peculiarities into account in system designs and avoid their shortcomings at the same time, the I/O bottleneck problem could be greatly mitigated.